EP0865031B1 - System for speech recognition of digits - Google Patents

Description

The invention relates to a system for speech recognition of digits a control arrangement for recognizing at least one sequence of digits and Output of the recognized digits of at least one sequence of digits.

Such a system is known, for example, from WO 95/06309 A1 and contains a remote control with a microphone and a circuit to form a modulated infrared signal. The voice inputs recorded by the microphone of a user are transmitted to a control device by means of an infrared signal, which converts the voice input into code words and an evaluation circuit for Formation of control commands, e.g. for a VCR or TV implements. With the help of individual voice inputs or commands, certain Functions can be carried out in the television or video recorder. For example can choose a channel, adjust the volume or play of a video tape can be stopped. It is also a time programming for the VCR described in which by entering the channel, date, start time and At the end of programming in a predetermined sequence is provided. Numbers are required for voice input. After entering of digits, for example for a channel or a time, from it Expenses formed to control the video device or control TV set. A comparison is made with saved ones Patterns instead. If a number is incorrectly recognized and assigned a wrong pattern incorrect control is carried out.

Furthermore, from the patent specification WO-A-89/04035 is an automatic Known speech recognition system that is part of a phone. Entering Telephone numbers are made by voice input from a user.

The invention is therefore based on the object of a system for speech recognition to create, in which the transmission of incorrectly recognized digits prevented becomes.

The invention is achieved by a system of the type mentioned in the introduction in that if at least one incorrectly identified digit of a first sequence of digits is provided, the control arrangement is provided for comparing a spoken second sequence of digits with the first sequence of digits,
that the control arrangement for determining connected digits of a part of the first sequence of digits, which has the most matches with the digits of the second sequence of digits, is provided if the number of digits of the second sequence of digits is less than the number of the first sequence of digits, and
that the control arrangement is provided to replace the digits of the determined part of the first sequence of digits with the non-matching digits of the second sequence of digits.

In the system according to the invention, verification of the voice inputs is carried out made by the user. The digits are specifically corrected here, that have not been understood. The speech recognition can according to the from the Publication "Hermann Ney, Volker Steinbiß, Xavier Aubert, Reinhold Haeb-Umbach: Progress in Large Vocabulary, Continuous Speech Recognition, in: H. Niemann, R. de Mori, G. Hanrieder: Progress and Prospects of Speech Research and Technology, 1994, pages 75 to 92 "are known methods. Here, linked strings of digits are created using hidden Markov models recognized. After an entered string of digits for the purpose of verification has been issued, the user has the option of the recognized sequence of digits to accept or reject and then certain numbers again enter. The digits are either output by the control arrangement by speech synthesis or by output of pre-recorded and saved Single digits. The tax arrangement must contain the digits 0 to 9 and certain tax inputs, such as. Understand "yes", "no" etc.

When a first sequence of digits has been recognized, the user is asked whether this has been understood correctly. If not, the user will asked to enter a new voice. Here he can complete one Enter new sequence of digits or just a partial sequence of digits. Subsequently the first sequence of digits and the newly entered second sequence of digits compared. The control arrangement then determines part of the first sequence of digits, which has the most numerical matches with the digits of the second sequence of digits having. The prerequisite is that the number of digits is the second String of digits is less than the number of the first string of digits. Then be the digits of the part of the first sequence of digits by the mismatched ones Digits of the second sequence of digits replaced.

Such a system can be part of a telephone, for example, in which the Telephone number is given via voice input. Furthermore, the invention System used in a value-added voice service (e.g. voice dialing in the network) become.

The system according to the invention has the advantage that only such digits with the corresponding context entered by the user in the event of a correction become. For example, only the number before and after one is used by the user incorrectly understood number announced again. This form of correction corresponds to that natural, familiar procedure and is also faster than that Enter the entire sequence of digits again. In addition, the prospect of success is at This type of correction is larger, because by entering a partial digit sequence the The risk of detection errors is lower.

The control arrangement determines the number of digits during the evaluation the first and second sequence of digits and determines which digits of all contiguous parts of the first sequence of digits with the second sequence of digits to match. If there are several parts or partial sequences of the first sequence of digits have the same number of matches, a substring of them is for the correction selected. Here, the first partial sequence determined can be several Substrings with a matching number of digits with the second Partial sequence can be selected.

Furthermore, the control arrangement for marking at least one replaced digit in the first partial sequence by a digit of the second partial sequence and for outputting a marked digit provided with an emphasis. The other digits too pronounced with a certain accent. The control arrangement specifies digits an odd position of the digit sequence with an increasing and digits at one straight and at the last position of the digit sequence with a falling accent. This form of natural output with pairwise prosody becomes a enables better recording of the digits. The output of a corrected digit with emphasis (contrasting accent) allows easier success monitoring.

After evaluating a first and a second sequence of digits, the control arrangement forms with the output of a corrected first sequence of digits to the User whether the number sequence is recognized correctly.

The invention also relates to a method for speech recognition of digits, in which at least one sequence of digits is recognized and the recognized digits of at least one sequence of digits are output. It is provided
that in the case of at least one incorrectly recognized digit of a first digit sequence, a spoken second digit sequence is compared with the first digit sequence,
that contiguous digits of a part of the first sequence of digits which have the most matches with the digits of the second sequence of digits are determined if the number of digits of the second sequence of digits is less than the number of the first sequence of digits, and
that the digits of the determined part of the first sequence of digits are replaced by the non-matching digits of the second sequence of digits.

Fig. 1

ein System zur Spracherkennung und

Fig. 2 bis 4

Ablaufdiagramme zur Erläuterung der Verfahren zur Spracherkennung.

Exemplary embodiments of the invention are explained in more detail below with reference to the figures. Show it:

Fig. 1

a system for speech recognition and

2 to 4

Flow diagrams for explaining the methods of speech recognition.

1 shows an exemplary embodiment of a system for speech recognition by Numbers shown, which is a microphone 1, two amplifiers 2 and 3, one Speech recognition arrangement 4, an evaluation circuit 5 and a loudspeaker 6 contains. The speech recognition arrangement 4 and the evaluation circuit 5 form one Control arrangement 33. Voice inputs of a user are made via the microphone 1 entered. The system receives special voice inputs, namely certain ones Numeric sequences (e.g. "3 8 7 4 2 1 6") and control inputs via the amplifier 2 be delivered to the speech recognition arrangement 4. The speech recognition arrangement 4 can, for example, from a signal processor with appropriate Peripheral circuits exist, the program running in operation one Speech recognition enabled. Such programs are known and the Methods underlying the program can e.g. the document "Hermann Ney, Volker Steinbiß, Xavier Aubert, Reinhold Haeb-Umbach: Progress in Large Vocabulary, Continuous Speech Recognition, in: H. Niemann, R. De Mori, G. Hanrieder: Progress and Prospects of Speech Research and Technology, 1994, Pages 75 to 92 ". The user entered and Recognized sequences of digits are used as code words (e.g. in ASCII code) Evaluation circuit 5 given. In the evaluation circuit 5 is a voice output arrangement included, which forms speech outputs from the recognized digits. Here, the speech output arrangement can either be from a speech synthesizer exist, which delivers the synthesized digits to amplifier 3, or Narrator takes stored from a memory Speech segments of a speaker, which are also fed to the amplifier 3.

The speech output is then via the speaker 6 to the user Check announced. Here, the evaluation circuit also determines Announcements or phrases output, such as: "Is the number sequence correct has been understood? ". The user can then use a misunderstood Make corrections to the digit or sequence of digits.

The evaluation circuit 5 also contains, for example, a microprocessor corresponding peripheral circuits, which software modules for processing the recognized control inputs and digits and for controlling the voice output arrangement having. 2 schematically shows the flow diagram of the main process shown to recognize digits. After a string of digits from the speech recognition device 4 what has been received by the abbreviated Notation ERK_ZN in a block 7 of FIG. 2 is indicated for the An analysis and accent marking (AN_AK, Block 8) carried out. There are digits in a sequence of digits with an odd number Position with a "b" and digits of the digit sequence with a straight position with marked with an "e". The last digit in the string of characters will be regardless of whether it is is an even or odd position of the digit sequence, marked with an "e". The means that the number is in the first, third, fifth position etc. of the number sequence with a "b" and the digit in the second, fourth, sixth position etc. the Sequence of digits are marked with an "e". Hereby for the output created a pairwise prosody. For example, the number sequence "3 8 7 4 2 1 6 "marked as follows:" 3b 8e 7b 4e 2b 1e 6e ".

The next step after block 8 is carried out in the flow chart according to FIG a block 9 (AU_KO) specified. This marks the output of the recognized Sequence of digits and the question of whether the sequence of digits was recognized correctly. In the Output of the number sequence uses the speech output arrangement of the evaluation circuit 5 two phrase variants. A digit will either increase or issued with a falling accent. For the digits marked with "b", phrases with increasing accent are used and for the digits that start with "e" phrases with a falling accent are used. This gives a pairwise prosody course, which is the natural Speaking behavior of people.

After the user's response to the system question has been recognized (Block 10, ERK_A), a query block 11 (OK?) Checks how the Answer is. If the user answered "Yes", the number sequence is recognized and the entry is completed. The recognized sequence of digits can then be used to Further processing can be used. If the answer is "no", the system asks the user for a correction, which is indicated by the abbreviation AU_FR in a block 12 is shown. The user can then create a completely new one Enter a sequence of digits or a partial digit sequence. The subsequent one Speech recognition process and analysis of the newly entered sequence of digits indicated by the abbreviation ERK_ZK in a block 13. To the speech recognition and analysis in block 13 is the one already described above further analysis and accent marking (block 8) carried out.

The analysis indicated by block 13 is shown in FIG 3 and 4 explained in more detail. The start of the analysis in FIG. 3 is identified by ST. First it is checked whether the length L (Z1) of the old sequence of digits Z1 is smaller than the length L (Z2) of the new digit sequence Z2 (block 14: L (Z1) <L (Z2)). is if this is not the case, the old sequence Z1 is replaced by the new sequence Z2 replaced what is indicated in a block 15 by Z1 → Z2. This then ends the analysis. However, if the digit sequence Z1 is greater than or equal to the new one Numeric sequence Z2, as listed in a block 16, the variables m, mT and mS set to zero (m = 0, mT = 0, mS = 0).

The following is the description of the part of the flow chart which finds out which part of the old number sequence Z1 is most similar to the new number sequence Z2.

At the beginning of a first loop, it is first checked (block 17) whether the Sequence of digits Z2 has been compared with each part of the sequence of digits of Z1. It it is checked in block 17 whether the value of the variable m is less than or equal to that The difference in the length of the digit sequences of Z1 and Z2 is: m ≤ L (Z1) - L (Z2). For example, if the digit sequence Z1 the digits "3 8 7 4 2 1 6" and the Sequence of digits Z2 has the digits "7 5 2", the length of the digit sequence is Z1 equal to 7 and the length of the digit sequence Z2 equal to 3. So the first loop be run through a total of five times, since a total of five parts of the digit sequence Z1 ("3 8 7", "8 7 4", "7 4 2", "4 2 1" and "2 1 6") with the digit sequence Z2 be compared. If the comparison in block 17 shows that the value of the variable m is greater than the difference in the length of the digit sequences of Z1 and Z2, the first is Loop ends and a transition is made to a second loop Flow chart is shown in Fig. 4. The transition to the second loop gives the Mark "A" (block 18).

If the comparison m L L (Z1) - L (Z2)) is true, as in a block 19 shown, two further variables n and t set to zero. The variable n gives in each case the position of a digit in the digit sequence Z2 and the variable t the number of Matching digits between the part of the part to be compared Sequence of digits Z1 and the sequence of digits Z2. The subsequent query block 20 is as well as blocks 21, 22 and 23 part of a sub-loop. In block 20 checks whether the value of the variable n is less than the length of the digit sequence Z2 is. If this is the case, query block 21 asks whether the digit is at the position m + n of the digit sequence Z1 is equal to the digit at position n of the digit sequence Z2 (Z1 (m + n) = Z2 (n)). If this query is answered in the affirmative, the variable t is incremented (Block 22). In the negative case, just as after the processing of block 22 jumped to block 23. Block 23 gives the incrementing of variable n on. The further processing is then continued in block 20.

Occurs when the value of the variable n is greater than or equal to the length of the Numeric sequence is Z2 (block 20), the further processing in a query block 24 continued. Here it is checked whether the value of the variable t is greater than that Value of the variable mT is. If this is the case, the variable mT becomes t and the Variable mS set equal to m (block 25). The variable mS gives the part of the Sequence of digits Z1, which matches most of the digits with the sequence of digits Z2 has. The variable mT contains the number of matches for this. In the next step after a negative query result of block 24 or after setting the variables mT and mS in block 25, the variable m is incremented, as shown in block 26. This completes the first loop, in a part of the digit sequence Z1 is determined, which with the digit sequence Z2 on best matches. In the example above, the part is with the digits "7 4 2" in the digit sequence Z1 with the digits "3 8 7 4 2 16" in the digit sequence Z2 most similar with the digits "7 5 2".

The second loop, which is shown in the flow diagram of FIG. 4, marks the Digits of the part of the digit sequence Z1 that differ from the digits of the digit sequence Z2. 4 begins with the mark "A" in the Block 27. Before the start of the second loop, the variable n is set to zero set, which shows a block 28. This variable n denotes the position of a Number in the string Z2. The second loop consists of the query blocks 29 and 30 and further blocks 31 and 32. In the query block 29 it is checked whether the value of the variable n is less than the length of the digit sequence Z2 (n <L (Z2)). If this is not the case, the analysis is ended. Otherwise it is checked whether a digit of the part of the digit sequence Z1 equal to the assigned digit of the Sequence of digits is Z2 (block 30). The mathematical expression for this is: Z1 (m + n) = Z2 (n). If the digit at position m + n of the digit sequence Z1 with the number at position n of the sequence of digits Z2 will be next went to block 32. Otherwise, if the digits do not match the number at the position n + m of the number sequence Z1 is replaced by the number at the Position n of the digit sequence Z2 replaced. This case is in block 31 by the Expression Z1 (m + n) → Z2 (n) marked. In addition, the replaced digit is marked with marked with an "a". This identification is in block 31 by the Expression a Z1 (m + n) specified. The next step is incrementing of the variable n, as shown in block 32. The further processing then takes place again at query block 29.

After the analysis has been performed in block 13 (FIG. 2), the new one Numeric sequence Z1 delivered to block 8. For example, the new sequence of digits is Z1 with the digits "3 8 7 5 2 16" from the old sequence of digits Z1 with the digits "3 8 7 4 2 16 "and the sequence of digits Z2 with the digits" 7 5 2 " the number "4" has been replaced by the number "5". Block 8 also receives from block 13 the marking of the replaced digit or digits with the Letter "a". The block is then, as described above, in block 8 of the digits with the letters "b" and "e". For the marked A corresponding voice output is generated in block 9. A Digit that is marked with "b" will appear with increasing accent and a digit that marked with "e" is output with a falling accent. A number with the letter "a" is highlighted with an emphasis, to notify the user of the change made. The marking of the new number sequence Z1 for the example is as follows: "3b 8e 7b a5e 2b 1e 6e ".

The loop shown in FIG. 2 in the flow chart from blocks 8 to 13 is run through until the user accepts the result.

It should also be mentioned that the control arrangement 33 is also designed as a computer system can be, which the functions of the speech recognition arrangement 4 and the Evaluation circuit 5 executes.

Claims (6)

1. A speech recognition system for digits, comprising a control device (33) for recognizing at least one spoken digit sequence and for producing the recognized digits of at least one digit sequence, characterized in that, when there is at least one erroneously recognized digit of a spoken first digit sequence, the control device (33) is provided for comparing a spoken second digit sequence with the first digit sequence, in that the control device (33) is provided for determining correlating digits of a part of the first digit sequence which sequence has the most matches with the digits of the second digit sequence when the number of digits of the second digit sequence is smaller than the number of digits of the first digit sequence, and in that the control device (33) is provided for substituting the non-matching digits of the second digit sequence for the digits of the determined part of the first digit sequence.
2. A system as claimed in claim 1, characterized in that the control device (33) is used for

   determining the number of digits of the first and second digit sequence,

   determining the match of the digits of all the correlating parts of the first digit sequence with the second digit sequence,

   selecting a part of the first digit sequence having the most matches if more parts of the first digit sequence have the same number of matching digits.
3. A system as claimed in claim 1, characterized in that the control device (33) is used for marking at least a substituted digit in the first digit sequence by a digit of the second digit sequence and for producing a marked digit with a sound.
4. A system as claimed in claim 1, characterized in that when a corrected first digit sequence is produced and after the evaluation of a first and a second digit sequence, the control device (33) is used for forming a question to the user whether the digit sequence has been recognized correctly.
5. A system as claimed in claim 1, characterized in that the control device (33) is used for producing the digits on an odd position of the digit sequence with rising stress and digits on an even and on the last position of the digit sequence with falling stress.
6. A speech recognition method of digits in which at least one spoken digit sequence is recognized and the recognized digits of at least one digit sequence are produced, characterized in that in the event of at least one erroneously recognized digit of a spoken first digit sequence, a spoken second digit sequence is compared with the first digit sequence, in that correlating digits of a part of the first digit sequence which sequence has the most matches with the digits of the second digit sequence are determined when the number of digits of the second digit sequence is smaller than the number of the first digit sequence, and in that the digits of the determined part of the first digit sequence are replaced by the non-matching digits of the second digit sequence.

Images